Google Taught Robots to Play Football

Researchers at DeepMind, a division of Google, have demonstrated success in teaching robots to perform motor actions. The team trained small bipedal bots to play football. The researchers explained how they managed to achieve this and why things will only get better.

The company trained the robots to play using deep reinforcement learning – that is, the bots were trained to process unstructured data. For example, this is not a set of numbers from 1 to 100, but completely unpredictable data, for example, a rapidly changing gaming environment on the playing field.

Robots act in such a way as to achieve a result that was labeled as positive at the time of training: for example, getting up after falling is a positive result. The difference between the achievement and other similar ones is that the method was applied to the physical coordination of robots in space.

Initially, the “subjects” were trained in a computer simulation with a concentration on two tasks: getting up after falling and scoring a goal. The researchers then combined these skills into a one-on-one soccer game. Playing more and more, the robots improved their skills to the point they showed in the video.

The number of motor actions included kicking, defense, reaction to the opponent’s actions and “aimed shooting” at the goal. The researchers also mention that robots exhibit non-standard behavior that is difficult to program in advance. But the training was carried out in a simulation, and in the future the company plans to begin “training” on the playing field to further increase the adaptability of behavior.